Estructuración de un algoritmo basado en deep learning para entrenamiento de robots asistentes en reconocimiento de objetos para plataformas multiherramienta
Fecha
Autores
Autor corporativo
Título de la revista
ISSN de la revista
Título del volumen
Editor
Compartir
Altmetric
Resumen
This document presents the development of an algorithm oriented to assistive robotics schemes working in multi-tool environments. This refers to an anthropomorphic robot that develops in a work area shared with a person, who will be assisted in tasks such as delivery of tools. For this, the robot must identify which tool they want to take within a group of tools, which corresponds to a work of pattern recognition. Each tool has particular characteristics that must be learned through a recognition algorithm. In this work, a difficulty in this recognition is exposed, that has not yet been addressed in similar works and that arose from the own developments of assistive robotics previously carried out. When looking to recognize a tool within a group, the pattern recognition system must learn the characteristics that each tool exhibits. Typically, this task is done by capturing the image of the tool group by means of a camera, from a given position, learning tool by tool. Once the tool is recognized and located spatially, trajectory planning algorithms are used, which by means of the robot's kinematics, allow the movement of its end effector towards the tool, but if the case arises that a user interrupts this path in the work area, the robot must look for a solution. Nowadays, the most used method has been to stop the robot and wait for the interruption to end, when seeking to improve this solution, where the robot is capable of generating evasion of what is obstructing it, a new trajectory must be sought, from the new point where it is located (displaced robot) to the tool. This is where the problem of recognition is presented, from the new position the information of the tool must be captured to generate the new trajectory, where, when changing the capture point by proximity or distance, the tool presents more or less characteristics, which varies the degree of recognition from the point of initial learning, making it difficult to recognize and confusing the tools present in the scene. This task from static points has been worked extensively, but from this dynamic perspective it does not yet present solutions, which are necessary to improve human-machine interaction, just as a human being does, that changes the trajectory to its destination when an obstacle is detected.